Context-aware rule learning from smartphone data: survey, challenges and future directions

A static segmentation is easy to understand and can be useful to analyze population behavior comparing across the mobile phone users. In order to generate segments, recently, most of the researchers (shown in Table 1) take into account only the temporal coverage (24-h-a-day) and statically segment time into arbitrary categories (e.g., morning) or periods (e.g., 1 h). Such static segmentation of time mainly focuses on time intervals. According to [45], there are mainly two types of time intervals: one is equal and another one is unequal. For instance, four different time segments, i.e., morning [6:00–12:00], afternoon [12:00–18:00], evening [18:00–24:00] and night [0:00–6:00] can be an example of equal interval based segmentation because of their same interval length. On the other hand, another four time slots such as morning [6:00–12:00], afternoon [12:00–16:00], evening [16:00–20:00] and night [20:00–24:00 and 0:00–6:00] can be an example of unequal interval based segmentation. For this example, different lengths of time interval are used to do the segmentation. In Table 1, we have summarized a number of works that use static segmentation considering either equal or unequal time interval in various purposes.

Although, various time intervals and corresponding segmentation summarized in Table 1 are used in different purposes, these approaches take into account a fixed number of segments for all users. However, while performing such segmentation users’ behavioral evidence that differs from user-to-user over time in the real world, is not taken into account. Thus, these static generation of segments may not suitable for producing high confidence temporal rules for individual smartphone users. For instance, \(N_1\) number of segments might give meaningful results for one case, while \(N_2\) number of segments could give better results for another case, where \(N_1 \ne N_2\). Therefore, a dynamic segmentation of time rather than statically generation could be able to reflect individuals’ behavioral evidence over time and can play a role to produce high confidence rules according to their usage records.

As discussed above, a segmentation technique that generates variable number of segments would be more meaningful to model users’ behavior. Thus, dynamic segmentation technique rather than static segmentation can be used in order to achieve the goal. In a dynamic segmentation, the number of segments are not fixed and predefined; may change depending on their behavioral characteristics, patterns or preferences. Several dynamic segmentation techniques in terms of generating variable number of segments exist for modeling users’ behavioral activities in temporal contexts. A number of authors simply take into account a single parameter, e.g., interval length or base period, to generate the segments. The number of time segments varies according to this period. If \(T_{max}\) represents the whole time period of 24-h-a-day and BP is a base period, then the number of segments will be \(T_{max} / BP\) [10]. If the base period increases, the number of time segments decreases and vice-versa. For instance, if the base period is 5 min, then the number of segments will be the division result of 24-h-a-day and 5. In this example, a base period, e.g., 5 min, is assumed as the finest granularity to distinguish day-to-day activities of an individual. If the base period incremented to 15 min, then the number of segments decreases, where 15 min can be assumed as the finest granularity. Thus the number of segments varies based on the base time period. Similarly, individuals’ calendar schedules and corresponding time boundaries can also be used to determine variable length of time segments, in order to model users’ behavior in temporal context, which may vary according to users’ preferences [59]. For instance, one user may have a particular event between 1 and 2 p.m., while another may have in another time boundary between 1:30 and 2:30 p.m.. Thus, the time segmentation varies according to their daily life activities scheduled in their personal calendars. Similarly, multiple thresholds, sliding window, data shape based approaches are used in several applications, shown in Table 2. In addition to these approaches, a number of authors use machine learning techniques such as clustering, genetic algorithm etc. In Table 2, we have summarized a number of works that use such type of dynamic segmentation techniques in various purposes.

Clustering highlighted in Table 2 is one of the important machine learning techniques in forming large time segments where certain user behavior patterns are taken into account. Usually, clustering algorithms are designed with certain assumptions and favor certain type of problems. In this sense, it is not accurate to say ‘best’ in the context of clustering algorithms; it depends on specific application [75]. Among the clustering algorithms the K-means algorithm is the best-known squared error-based clustering algorithm [76]. However, this algorithm needs to specify the initial partitions and fixed number of clusters K. The convergence centroids also vary with different initial points. Sometimes this algorithm is influenced by outliers because of mean value calculation. More importantly, the characteristic of this algorithm might not be directly applicable for the purpose of learning context-aware rules. The reason is that users' behave differently in different contexts, which also may vary from user-to-user in the real world. Thus, it's difficult to assume a number of clusters K to capture their diverse behaviors effectively. Another similar K-medoids method [77] is more robust than K-means algorithm in the presence of outliers because a medoid is less influenced by outliers than a mean. Though it minimizes the outlier problem but the other characteristic mismatches exist between K-means and the problem of time-series modeling.

As the size and number of time segments depend on the user’s behavior and it differs from user-to-user, a bottom-up hierarchical data processing can help to make behavioral clusters. Existing hierarchical algorithms are mainly classified as agglomerative methods and device methods. However, the device clustering method is not commonly used in practice [75]. The simplest and most popular agglomerative clustering is single linkage [78] and complete linkage [79]. Another method, nearest neighbor [75], is also similar to the single linkage agglomerative clustering algorithm. All these hierarchical algorithms use a proximity matrix which is generated by computing the distance between a new cluster and other clusters. Then according to the matrix value these algorithms successively merge the clusters until the desired cluster structure is obtained. However, it is not possible to predict the level at which the merging is best according to a proximity matrix because of the variations in users’ behavior. Thus applying such clustering techniques could generate the segments according to users’ behavioral patterns available in time-series. Similarly, genetic algorithm based approaches shown in Table 2 also produce dynamic segments.

In a summary, we can conclude that time-series modeling in terms of both static segmentation and dynamic segmentation approaches discussed above, are able to generate various time segments that can be used in different purposes. However, the above time-series modeling approaches do not necessarily map to the patterns of individuals’ usage according to their preferences, which is based on users’ diverse behaviors over time-of-the-week and may vary from user to user. A machine learning based behavior-oriented dynamic time-series modeling technique by taking into account such patterns, could be significant in order to effectively use temporal context as the basis for discovering rules capturing smartphone usage behavior.

Association rule learning algorithm discovers association rules that satisfy the predefined minimum support and confidence constraints from a given dataset [80]. Many association rule learning algorithms have been proposed in the data mining literature, such as logic based [82], frequent pattern based [80, 83, 84], tree-based [85] etc. Association rule learning technique is well defined in terms of rule’s performance, e.g., accuracy, and flexibility as it has the own parameter support and confidence [86]. A number of researchers [7‐9] have used association rule learning technique (e.g., Apriori) [80] to mine rules capturing mobile phone users’ behavior. However, the existing association rule learning techniques might not be suitable for discovering users’ behavioral rules because of several reasons. In the following, we summarize the drawbacks of association rules for discovering the behavioral rules of individual mobile phone users by taking into account multi-dimensional contexts.

Lacking in understanding the impact of contexts Different contexts in mobile phone data, such as temporal, spatial or social context, may have different impact or influence on the behavioral rules of individual mobile phone users. For instance, incoming phone calls from a significant person, e.g., mother, is always answered by an individual, even though she is in a meeting because of her family priority. In this case, the importance of social relationship between individuals (\(social \; relationship \rightarrow mother\)) in making behavioral decision, is higher than other relevant contexts such as time period, weekday or holiday, location, accompany with etc. However, the typical association rule learning technique implicitly assumes all the contexts in the datasets have the similar nature, and/or impact while discovering rules based on multi-dimensional contexts.

Redundancy Association rule learning technique, e.g., Apriori, discovers all the contextual associations in a given dataset, if it satisfies the user preference, specified as minimum support value and minimum confidence value. As a result, association rule learning technique produces a huge number of redundant rules as it does not take into account the usefulness of a particular context or corresponding patterns while producing the associations. For instance, it produces up to 83% redundant rules for a given dataset that makes the rule-set unnecessarily large [87]. Therefore, it is very difficult to determine the most interesting ones among the huge amount of rules generated. As a result, it makes the rule-based decision making process ineffective and more complex, which is not effective to build a context-aware intelligent system [88].

Computational complexity and high training time In order to produce rules, association rule learning technique takes huge amount of training time. For instance, in an experimental study in mobile phone domain, the authors observe a high running time spanning several hours when the association rule learning algorithm is used to discover user behavioral rules [8]. The main reason for taking high training time is that typical association techniques compute all the possible associations among contexts and are unable to filter the interesting rules that can be used to make effective decisions. As a result the unnecessary generation of patterns increases the computational complexity and training time.

In a summary, by taking into account the impact of contexts, redundancy problem while generating rules, and computational complexity, typical association rule learning techniques may not suitable to produce users’ behavioral rules in multi-dimensional contexts, for the purpose of building intelligent context-aware systems.

Classification is another technique to discover user behavioral rules from the datasets. Several classification algorithms exist with the ability of rule generation like ZeroR [89], OneR [90], RIDOR [89], RIPPER [91], PART [92], DTNB [93], Decision Trees [81, 94] etc. Among these techniques, decision tree is one of the most popular rule-based classification algorithms as it has several advantages, such as easier to interpret; the ability to handle high dimensional data; simplicity and speed; good accuracy; and the ability to generate human understandable classification rules [95, 96]. In particular, a number of authors [67, 97‐100] have used decision tree classification technique to discover rules capturing mobile phone users’ behavior for various purposes. However, the exisitng rule-based classification techniques might not be suitable to model users’ behavior because of several reasons. In the following, we summarize the drawbacks of rule-based classification techniques for discovering the behavioral rules of individual mobile phone users in multi-dimensional contexts.

Low reliability In general, reliability refers the quality of being trustworthy or of performing consistently well. A pattern or rule is called reliable, if the relationship described by the pattern occurs in a high percentage of applicable cases. According to Geng et al. [101], a classification rule will be reliable, if it gives high prediction accuracy, and an association rule will be reliable, if it has high confidence that is associated to the accuracy. However, the classification rules discovered by the typical rule-based classification techniques, e.g., decision trees, mostly have low reliability in many cases [7, 102]. According to Freitas et al. [86], a classification rule may not ensure a high accuracy in predictions. The reason is that it may has over-fitting problem and inductive bias, which decrease the prediction accuracy of a machine learning based model.

Lacking in flexibility Traditional rule-based classification techniques, e.g., decision trees, have no flexibility to set users’ preferences and consequently it makes rigid decision for a particular test case [81]. However, rigid decision in modeling user behavior might not be meaningful by considering real-world cases. The reason is that individuals’ preferences are not static in the real word; may vary from user-to-user [103]. For instance, one individual may want the phone call agent to decline the incoming calls where she did not answer the calls more than, say, 80% of the time in the past. For another person, this preference could be 95% of the time according to her preference. Thus considering flexibility in users' preferences while modeling their behavior could be another issue for making meaningful decisions in various context-aware test cases. 

Lacking in generalization Typically, generality measures the comprehensiveness of a pattern or rule, that is, the fraction of all the relevant records in the dataset that matches the pattern. According to Geng et al. [101], if a pattern characterizes more information in the relevant dataset, it tends to be more useful and interesting. Traditional classification techniques consider data-driven generalization while producing classification rules. Besides this, users’ behavior-oriented generalization might be interested for learning context-aware rules. For instance, users’ behavior might be similar for a collection of contexts and have exceptions only in few cases [6]. Thus, users’ behavior oriented generalization could give more precise results for modeling their usage behavior. The generalization not only simplifies the resultant machine learning based model but also minimizes the over-fitting problem and improves the prediction accuracy.

In a summary, by taking into account the reliability, flexibility, and generalization discussed above, typical classification rule learning techniques may not be suitable to produce users’ behavioral rules in multi-dimensional contexts, for the purpose of building intelligent context-aware systems.